<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0306">
  <Title>Word Alignment Baselines</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
8 Conclusion
</SectionTitle>
    <Paragraph position="0"> Several baseline alignment systems were presented. The individual scores of the different aligners give insight into the relative contributions of the features they exploit.</Paragraph>
    <Paragraph position="1"> Word length matching appears to be the least important feature, followed by character edit distance (attempting to match cognates), and geometric dotplot distances appear to contribute most strongly to alignment performance.</Paragraph>
    <Paragraph position="2"> The supervised probabilistic models perform poorly on their own, probably because of the unconstrained way in which they were trained and applied. When all features are combined in concert into a larger alignment system using the nearest neighbor rule, they perform better than individual aligners, but the question remains of what space should be used for modeling the points (distances versus binary decisions).</Paragraph>
  </Section>
class="xml-element"></Paper>